Mohamed Khayat

This repository provides a framework for fine-tuning Hugging Face transformer-based object detection models on custom datasets, featuring a comprehensive training loop with advanced techniques (gradient accumulation, FP16, EMA, etc.), a Streamlit inference UI, and tools for uploading models to Hugging Face Hub. It is demonstrated on a fruit object detection task with 16 models achieving mAP scores up to 0.68.

mini-llm-lab is a configurable framework for pretraining decoder-only language models from scratch, featuring custom implementations of model components, a two-stage warmup-stable-decay training schedule, bit-exact resume capability, and an integrated chat interface for interacting with trained checkpoints.

The repository implements the Show, Attend and Tell image captioning model with soft attention, allowing comparison of different encoders (VGG19, ResNet50, CLIP ViT-B/16) on Flickr8k. It reproduces the paper's architecture and reports BLEU scores, noting deviations such as freezing the encoder weights.

DIYNeuralNet is a from-scratch NumPy-based neural network library designed for educational purposes, featuring hand-derived forward and backward passes for layers, activations, losses, and optimizers, accompanied by detailed documentation and runnable notebooks.

This repository provides a flexible PyTorch-based training pipeline for image classification tasks, supporting various architectures via Hydra configuration and integrating Weights & Biases for experiment tracking. It includes custom CNN models, data loading utilities, training loops with mixed precision and Grad-CAM, and early stopping.

This repository provides an end-to-end machine learning application for estimating medical insurance premiums. It includes a CRISP-DM based training notebook, a FastAPI serving API with model inference, LIME explanations, customer segmentation, and a simple UI.